WEBVTT

00:00:00.240 --> 00:00:06.640
so at some point after robots take all of our jobs they'll eventually decide

00:00:04.560 --> 00:00:10.320
that the world just really doesn't need humans that badly and will do away with

00:00:08.960 --> 00:00:15.120
us all together but then haven't you ever wondered how

00:00:13.120 --> 00:00:19.279
this incredibly rosy vision of the future will come to pass

00:00:16.960 --> 00:00:24.960
well i can tell you that it'll probably involve neural networks now we tend to

00:00:21.840 --> 00:00:27.119
think of artificial intelligence as the

00:00:24.960 --> 00:00:32.480
capability of a computer to make sensible decisions on its own kind of

00:00:29.679 --> 00:00:36.880
like a human and as it turns out it's more human-like than you actually might

00:00:34.719 --> 00:00:42.320
think computer scientists and programmers have modeled the way that

00:00:39.680 --> 00:00:48.559
computers arrive at their decisions after our own brains at least to a point

00:00:45.840 --> 00:00:53.920
so inside your nervous system your neurons are arranged in such a way that

00:00:51.120 --> 00:00:59.039
after information is taken in one group of neurons will pass that information to

00:00:56.399 --> 00:01:02.559
the next group in a way that depends on what signals the previous group was

00:01:01.280 --> 00:01:07.840
sending okay so for example then if neurons in

00:01:05.119 --> 00:01:10.799
your muscles sense that they're working really hard because you're going out for

00:01:09.840 --> 00:01:16.240
a run those neurons send out signals through a

00:01:13.520 --> 00:01:21.759
specific pathway that ultimately results in your brain telling your diaphragm to

00:01:19.040 --> 00:01:25.119
breathe harder rather than you know send a message to your hair to grow

00:01:24.000 --> 00:01:30.560
faster and although the way that our neurons

00:01:27.439 --> 00:01:32.640
work is immensely more complicated than

00:01:30.560 --> 00:01:37.200
an artificial neural network made up of ones and zeros the neural networks used

00:01:35.040 --> 00:01:43.759
in machine learning are at least loosely based on the workings of your brain

00:01:41.040 --> 00:01:49.280
instead of neurons made up of dna and cytoplasm though a software neural

00:01:46.159 --> 00:01:52.320
network uses processing nodes stacked

00:01:49.280 --> 00:01:54.720
into many different layers so a node in

00:01:52.320 --> 00:01:58.560
the first layer which initially receives the data

00:01:55.759 --> 00:02:03.200
weights all of its incoming connections by multiplying each piece of data by a

00:02:01.360 --> 00:02:07.759
certain factor depending on the connection it then adds these numbers

00:02:05.600 --> 00:02:12.239
together and if it's above a certain threshold the node sends it along to a

00:02:09.679 --> 00:02:15.920
node in the next layer and then so on and so forth

00:02:13.680 --> 00:02:20.879
once the data gets to the end of the neural network the network has

00:02:17.680 --> 00:02:23.120
effectively made a decision based on

00:02:20.879 --> 00:02:26.800
weighting this combined and then recombined data across many of these

00:02:25.680 --> 00:02:31.120
layers now all of this probably sounds quite

00:02:29.680 --> 00:02:37.200
abstract but basically neural networks use thousands or

00:02:34.640 --> 00:02:42.000
millions of these nodes to turn those numbers into something useful

00:02:39.680 --> 00:02:46.239
of course even the most sophisticated neural network isn't going to know right

00:02:44.239 --> 00:02:51.519
off the bat how to differentiate a picture of a burrito from a body pillow

00:02:49.360 --> 00:02:54.480
just like a baby doesn't pop out of the womb already knowing how to speak the

00:02:53.040 --> 00:03:00.640
king's english so instead developers train a neural

00:02:57.840 --> 00:03:06.640
network by feeding it lots and lots of both relevant and irrelevant inputs so

00:03:04.159 --> 00:03:11.599
for something like a self-driving car it might be a ton of images including some

00:03:09.680 --> 00:03:16.400
of pickup trucks as the data gets fed into the network

00:03:13.760 --> 00:03:20.879
developers monitor its behavior and then adjust the weights on the processing

00:03:18.720 --> 00:03:26.080
nodes until the desired output is reached and the system can eventually

00:03:23.200 --> 00:03:31.200
quickly differentiate a truck from a bicycle or baby stroller or stationary

00:03:29.280 --> 00:03:36.560
object or what have you the whole idea is very general purpose

00:03:34.080 --> 00:03:40.000
so it can actually be adapted to lots of different situations

00:03:38.159 --> 00:03:44.319
other than the more well-known examples that we've already discussed like image

00:03:41.840 --> 00:03:48.640
recognition and autonomous vehicles neural networks have been used for

00:03:45.920 --> 00:03:52.799
everything from filtering spam out of email to training computers to play

00:03:50.879 --> 00:03:58.000
video games nearly as well as professional level human players

00:03:55.200 --> 00:04:02.319
of course training a computer to think is more complicated than getting it to

00:03:59.840 --> 00:04:06.720
run a bunch of pre-defined instructions and as such training a neural network

00:04:04.640 --> 00:04:11.439
can be quite time consuming and depending on the application requires

00:04:08.799 --> 00:04:15.599
some pretty powerful hardware in fact the computers in current self-driving

00:04:13.280 --> 00:04:21.680
cars are not only more powerful than your standard desktop pc they are

00:04:18.320 --> 00:04:23.520
notorious battery hogs but as processors

00:04:21.680 --> 00:04:29.440
become more efficient and our training methods become less tedious we fully

00:04:26.080 --> 00:04:32.240
expect that one day our cars and spam

00:04:29.440 --> 00:04:36.800
filters might be even smarter than we are so even if they do decide to turn on

00:04:34.800 --> 00:04:41.360
us take over the world and rule in tyranny at least we won't have to deal

00:04:39.040 --> 00:04:45.199
with scam emails from nigerian princes or terrible drivers anymore

00:04:43.440 --> 00:04:49.199
private internet access supports a variety of vpn protocols and types of

00:04:47.520 --> 00:04:53.440
encryption and authentication to allow you to dial in the exact level of

00:04:51.120 --> 00:04:58.479
privacy protection that you need they've got apps for Windows macOS Android iOS

00:04:56.240 --> 00:05:02.639
Linux and google chrome with support for other platforms coming you can connect

00:05:00.720 --> 00:05:07.360
up to five devices at the same time with a single account and their apps include

00:05:04.639 --> 00:05:12.080
features like dns leak protection ipv6 leak protection and an internet kill

00:05:09.680 --> 00:05:15.840
switch that will block all traffic if the vpn becomes disconnected

00:05:13.759 --> 00:05:20.000
unexpectedly so check it out today at the link in the video description

00:05:18.240 --> 00:05:23.120
so thanks for watching guys like dislike check out our other videos leave a

00:05:21.600 --> 00:05:28.400
comment if you have a suggestion for a future fast as possible episode and

00:05:25.520 --> 00:05:32.479
don't forget to subscribe to techwiki and ring the bell you look really

00:05:30.479 --> 00:05:35.280
closely make sure it's wrong ringed

00:05:33.680 --> 00:05:38.280
wrong right